A neural-network surrogate recorded large online speedups in two-dimensional numerical tests, but its accuracy deteriorated sharply in the high-contrast experiments. At lower contrast, mean total online speedup rose from 4.824-fold at n = 0 to 16.341-fold at n = 2. In high contrast, the neural solution's mean relative L2 error was 0.96005, 0.95584 and 0.94442 across those levels, while the reported modeling error rose from 0.27945 to 0.90271.
A shortcut for repeated local solves
The work is an arXiv version 2 preprint dated 4 September 2026. It proposes NN-MsHDG, a method in which a neural network predicts each coarse block's discrete Dirichlet-to-Neumann matrix, which represents its boundary-to-flux response, along with a source-induced flux vector from the local coefficient representation. The prediction replaces repeated local fine-scale operator construction during online assembly.
The comparison kept the numerical setup the same for standard MsHDG and NN-MsHDG. Both used the same fine and coarse discretizations, the same global skeleton coupling and a stabilization setting of 1 on every fine-grid face. The local Dirichlet-to-Neumann matrices and source vectors were the only components replaced by network predictions.
For each of two coefficient-contrast regimes, the two-dimensional training set contained 5,000 independent local reference-block realizations. The researchers used 80% for training and 20% for validation, then evaluated independent global realizations only after training. A separate one-dimensional verification used 1,200 local realizations for each coefficient resolution and contrast, split into 1,000 training and 200 validation samples, with global errors averaged over 25 independent realizations for each setting.
The gain was clearest at lower contrast
At lower contrast, the standard and neural solutions initially tracked one another closely. At n = 0, their mean relative L2 errors were 0.09782 and 0.09637; at n = 1, they were 0.06683 and 0.06534. At the richest tested trace level, n = 2, the standard error fell to 0.03525, while the neural error was 0.04516. The reported modeling error was 0.01351 at n = 0, 0.01338 at n = 1 and 0.03001 at n = 2.
The timing advantage increased as the trace level became richer. Mean assembly speedup rose from 8.397-fold at n = 0 to 15.811-fold at n = 1 and 30.530-fold at n = 2. Mean total online speedup followed the same pattern, reaching 16.341-fold at n = 2 from 4.824-fold at n = 0.
High contrast exposed the accuracy problem
The pattern changed in the high-contrast tests. Standard MsHDG's mean relative L2 error was 0.94864 at n = 0, 0.84478 at n = 1 and 0.43394 at n = 2. The neural method remained near 0.95, with errors of 0.96005, 0.95584 and 0.94442. Its reported modeling error increased sharply across the same levels, from 0.27945 to 0.69909 and then 0.90271.
Local diagnostics offered a more favorable picture, but not a guarantee of global accuracy. On held-out high-contrast reference-block data, mean local Dirichlet-to-Neumann action errors were 2.00 × 10−2 for smooth probes and 2.33 × 10−2 for random probes at n = 0. At n = 1, they were 9.74 × 10−3 and 1.06 × 10−2. These probe tests were much smaller than the global modeling discrepancy and did not establish that the assembled solution would be accurate.
The timing advantage persisted despite the accuracy warning. In high-contrast experiments, mean assembly speedups were 8.403-fold, 15.794-fold and 30.667-fold at n = 0, 1 and 2. Mean total online speedups were 4.805-fold, 8.670-fold and 16.764-fold.
A narrower result than the speed figures suggest
The one-dimensional check was more reassuring on accuracy than on speed. Its global modeling error stayed below 1% in all reported cases and did not grow as the number of coarse blocks increased. But measured total online speedup was only about 0.98 to 1.00, so that experiment functioned as a verification rather than an acceleration demonstration.
The reported timing figures cover the online stage only. Offline data generation and network training were excluded from the comparison. The supplied paper also reports a training-validation split and checkpoint selection, but no preregistration or formal power or sample-size rationale.
Paper data and sources
Original title: A Neural-network-based multiscale Hybridizable Discontinuous Galerkin method for solving PDEs in porous media
Authors: Tony Haines, Ke Shi
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text